ViT-FLSFO: A method for Cervical Image Classification

Abstract Cervical cancer ranks as the fourth most prevalent cancer among women worldwide. Early diagnosis facilitates timely intervention and treatment. Traditional colposcopy is a widely employed technique for evaluating cervical lesions. Recently, deep learning has been increasingly applied to the diagnosis of cervical diseases. However, conventional Vision Transformer (ViT) models face limitations in effectively extracting localized information, posing challenges in image recognition. This study introduces a novel model for classifying cervical images into normal, cervical intraepithelial neoplasia, and cancerous categories. The model incorporates a Feature-token Labeling Selection Module, Feature Fusion Module，and Overlapping Sampling Module to improve the ViT model (ViT-FLSFO). This module can realize the extraction of local feature areas and short-range interaction in the image, and enhance the edge capturing ability of the model and the proficiency of solving complex problems, thereby elevating prediction accuracy. Empirical results demonstrate that the ViT-FLSFO model achieves a detection accuracy of 91.88% in the cervical image dataset, a precision of 92.91%, a recall of 91.92% and an F1-Score of 91.99%, surpassing other advanced models. Consequently, this model holds significant potential for rapid auxiliary diagnosis in cervical imaging, contributing to the early detection and treatment of cervical cancer.